Predicting Electricity Consumption in Residential Building’s Using Deep Transfer Learning with Attention-LSTM

The speedy growth in urbanization has resulted to considerable growth in electricity consumption with about 40% consumed by building sectors. To reduce the gap for future demands, there is need for efficient load forecasting using novel paradigms. Attention-LSTM focuses on input features that are crucial to a prediction model and have theoretically proven to be better prediction approach with better accuracy. This strategy can extract the critical feature sequences efficiently and lower the impact of the unused feature sequences that may results from the attention weight. Additionally. Occupancy behavior have huge impact to energy consumption in buildings. However, most existing study ignores the occupancy profile uncertainty which can led to a substantial performance gap between the actual and predicted energy. Therefore, this research work focuses on improving the prediction performance using attention base LSTM transfer learning mechanism on a novel Jos electricity distribution company (JEDC) dataset. Upon evaluation with different baseline models, the proposed model outperforms the baseline models with an average MAPE of 0.0116 and CV-RMSE of 0.0113. These results opine that transfer learning base on Attention-LSTM method can efficiently select the critical predictive features and can helps track uncertainties to building electricity consumption with accuracy improvement to some extent.